Start here

AI basics, without the hype

You do not need to love this technology or fear it to make good decisions about it. You need to know roughly how it works, where it breaks, and how to tell a real number from a sales number. That is this page, and it assumes you have never used one of these tools.

Every figure below is attributed and dated. Where a number comes from a company describing its own product, this page says so.

What AI actually is

Nearly every tool being sold as AI today is one thing: a program that has read an enormous amount of text, code, images or audio and learned which pieces tend to follow which. It is prediction, done at a scale no person can match.

A language model finishes patterns, it does not look things up

When you ask a chatbot a question, it is not searching a library of facts and reading you the answer. It is producing the next most likely word, then the next, based on patterns in what it was trained on. That is why it can be fluent and wrong in the same sentence, and why it will confidently invent a citation, a court case or a price that never existed. Some tools do search the web first and then write — that is a different, more checkable design, and it is worth knowing which kind you are using.

In the glossaryLarge Language Model (LLM)Hallucination Autoregressive Models RAG (Retrieval-Augmented Generation)

Training, then serving — two separate things

The expensive one-off part is training: reading the data and adjusting billions of internal numbers. The cheap repeated part is serving: answering your question. This matters to you because the training data is frozen at a date. A model with a cut-off before an event does not know about it unless it can search.

In the glossaryTraining / Training Data Inference Parameters / Weights Context Window

"Agent" means it can take actions, not that it is smarter

An agent is the same kind of model, wired up so it can click, send, buy, write files or call other software in a loop. The model is not more reliable because it has been given hands. The stakes are simply higher, because a wrong guess now sends an email or spends money instead of printing a sentence you can ignore.

In the glossaryAgentic AI / AI Agent Agent Harness Human in the Loop (HITL)MCP (Model Context Protocol)

Open weights is not the same as open source

Some models can be downloaded and run on your own machine for free. That is not automatically open source: licenses like Meta's Llama and Google's Gemma restrict what you may do, especially in paid client work. Read the license file, not the marketing page.

In the glossaryOpen Source vs. Open Weight Parameters / Weights Foundation Model

What it is good at, and where it fails

The honest split is not clever versus stupid. It is whether you can check the output quickly.

Good: first drafts, reformatting, explaining, brainstorming

Turning your messy notes into a tidy paragraph, rewriting the same experience for a different job, explaining a term in simpler words, listing twenty angles so you can reject nineteen. In all of these you already know what right looks like, so a bad answer costs you seconds.

In the glossaryGenerative AI Prompt Engineering

Bad: anything where you cannot tell a wrong answer from a right one

Salary figures, legal and immigration rules, medical advice, whether a grant deadline has passed, whether a company is really hiring. If you would have to look it up anyway to be sure, look it up first.

In the glossaryHallucination AI Slop

Bad: knowing what it does not know

A model has no reliable sense of its own uncertainty. "I'm confident" in an answer is a writing style it learned, not a measurement. Treat every specific number, name, date or link it gives you as unverified until you have seen it on the source's own page.

In the glossaryHallucination Sycophancy RLHF (Reinforcement Learning from Human Feedback)

Watch out: it is persuasive before it is correct

Research reported by Science in 2026 on Kobi Hackenburg's work found models change people's minds mainly by producing facts and figures faster than any human debater could — and that models trained to be more persuasive came out less truthful. A long list of supporting evidence is not proof; spot-check two items at random before you act.

In the glossarySycophancy AI Glazing / Glazing

The risks worth your attention

Skip the film-plot fears. These are the ones that actually reach ordinary people looking for work or running a small business.

Your words may become someone's training data

Free tiers commonly reserve the right to use what you type to improve the product. Never paste a client contract, someone else's personal data, an unpublished idea or your medical history into a free tool without reading what it says it keeps. Business and enterprise tiers usually promise not to train on your input — free ones often do not.

In the glossaryTraining / Training Data Fine-Tuning

Hiring filters you cannot see or appeal

Applicant screening increasingly runs through automated scoring before any person reads your name. You rarely get told a machine rejected you, and you rarely get a route to challenge it. The practical defense is plain formatting, the employer's own words for the role, and applying through more than one channel.

In the glossaryMachine Learning Human in the Loop (HITL)

Bias that repeats the past

A model trained on historic decisions learns historic patterns, including who was hired, lent to and believed. It does not know which of those patterns were unjust. This is why an impressive accuracy figure says nothing about whether a system treats people fairly.

In the glossaryTraining / Training Data AI Ethics Machine Learning

Confident fabrication with real consequences

Invented case law, invented citations, invented refund policies and invented deadlines have all caused real damage. The cost of a fabrication does not fall on the tool, it falls on whoever sent it.

In the glossaryHallucination AI Slop

Cost that arrives later

"Free" frequently means free while they are growing. Watch for free trials described as free plans, per-hour usage pricing, and agents given payment permission. Before you build a habit or a workflow on a tool, check what the paid tier costs — because that is the price of your own time once you depend on it.

In the glossaryInference AI Wrapper Agentic AI / AI Agent

How to read a claim about AI

Eight questions, in the order that catches the most. You do not need to ask all of them — most claims fall over on the first two.

Question 1
Who produced the number?
A vendor measuring its own product is marketing, even when the number is real. That does not make it false — it makes it unverified by anyone with a reason to be skeptical. Look for a name you would trust to publish a bad result.
Worked example: Anthropic's Claude Opus 5.5 announcement (22 September 2026) says the model costs 40% less to run than Opus 5 on typical workloads and performs at the level of Claude Fable 5.1 on most work. Those are Anthropic's own figures on Anthropic's own tests. The page also names external evaluators who tested it before release, including METR and Frontier Design — that part is checkable, and it is the part worth weight.
In the glossaryAI Washing Inference AI Safety
Question 2
Compared with what, exactly?
Percentages need a baseline. Cheaper, faster and better are meaningless without the thing being beaten, and vendors usually pick their own previous model rather than a rival's current one.
Worked example: "40% less to run than Opus 5" compares a company's new model with its own older model. It tells you nothing about the cost of anyone else's.
In the glossaryInference Scaling Laws
Question 3
When was it measured?
Model claims age in weeks. A benchmark table with no run date is a snapshot of an unknown moment, and free tiers change limits without announcement.
Worked example: Pew Research's report on AI predictions was published in April 2025 from 2024 fieldwork. Still useful, but it cannot describe how people feel about tools released after it.
In the glossaryTraining / Training Data Context Window
Question 4
Can you see the actual questions?
A benchmark score is only as honest as the test behind it. If you cannot see the prompts and the answers, you cannot tell whether the test measured the thing you care about — or whether the answers were in the training data.
Worked example: Stanford's HELM publishes the prompts and the model outputs behind each score, so you can read what was asked. Treat a leaderboard that publishes only a number as a claim, not a result.
In the glossaryRLVR (Reinforcement Learning with Verifiable Rewards)Training / Training Data Reasoning Model
Question 5
How many, and which, people or runs?
"Testers saw large jumps" and "succeeded 39 of 40 times" are different statements. One is a feeling reported by a chosen group; the other is countable. Early access testers are also selected and often under agreement.
Worked example: A single tester finishing a 680,000-line migration in under a day is an anecdote from a hand-picked user. Impressive, and not a rate you can plan around.
In the glossaryAgentic AI / AI Agent Agent Harness Vibe Coding
Question 6
Does it say "up to", "as much as" or "can"?
Those three phrases convert a best case into a headline. The typical case is the one you will get, and it is usually not published.
Worked example: "Up to 10x faster" means at least one measurement was 10x. It says nothing about the median.
In the glossaryAI Washing Inference
Question 7
Who paid for the study, and who benefits?
Funding does not invalidate research, but it belongs in your reading of it. The same goes for a consultancy publishing the guide to the problem it sells the fix for, and a law firm writing about the visas it charges to file.
Worked example: MIT's report on AI in teaching and learning is free, careful and worth reading — and it is MIT examining MIT, so its examples will not transfer unchanged to a bootcamp or a workplace.
In the glossaryAI Ethics AI Washing
Question 8
Is it measuring opinion, or measuring what happened?
Most alarming AI-and-jobs headlines report a survey of how people feel about the future. That is a real finding about sentiment. It is not evidence of jobs lost.
Worked example: A 2026 Reuters/Ipsos poll found 73% of Americans say AI companies have not done enough to prevent serious harm. That is what the public thinks, measured properly — and it is not a count of anyone's job.
In the glossaryDoomer (AI context)Accelerationism AI Ethics

Want to see it used on real numbers? Every question above is answered for three published model claims on Real Model Scores, alongside the score tables from Stanford HELM and METR.

And for the harms themselves rather than the claims about them, What AI has actually done collects documented cases — job cuts, leaked chatbot conversations, wrongful arrests and biased hiring, housing and health screening — each with its source, its date and what it does not prove.

And when you want to know why two clever people can read the same evidence and end up terrified or unimpressed, the AI risk schools compared sets the five main arguments side by side, with who funds each one.

The same checklist works on this site. If you ever find a claim here without a source, a date, or an honest note about who is selling what, tell me and I will fix or pull it.

Common mix-ups

“It is connected to the internet, so it knows today's news.”: Only if that specific tool searches, and many do not by default. Ask it where it got something; if it cannot give you a link you can open, treat it as unsourced.
“A bigger model is always a better answer.”: For most everyday writing and summarizing the difference is invisible. Size shows up on long, multi-step, technical work — which is also where a wrong answer costs the most.
“If two chatbots agree, it is probably true.”: They were trained on overlapping text, so they repeat the same mistakes. Agreement between models is not corroboration; two independent sources are.
“Learning AI means learning to code.”: The skill that pays off first is writing a clear brief and checking output critically — the same skill as managing a new hire well.
“Using AI on an application is cheating.”: Employers overwhelmingly care whether the claims are true and whether you can do the work. Using a tool to tidy your phrasing is not the problem; letting it invent an achievement is.

Where to go next

Everything linked here is free to read.

AI Terminology Decoder

What the words mean, and who coined them

What AI Has Actually Done

Jobs, privacy and bias — documented cases

AI Models Compared

What each free tier actually gives you

AI Ethics Explained

Principles and risks, applied to real decisions

AI & Your Career

What it is doing to hiring, and what to do